Low-light Image Enhancement Based On Retinex Theory

The application of image processing technology is extremely common in current society,which shows its noticeable strength in military and civil fields.Image processing can be divided into multiple technologies: image inpainting,noise reduction and enhancement,etc.,the image enhancement technology can make the original blurry or poorly recognizable image clear.In real life,many pictures are unsatisfactory due to different reasons such as insufficient illumination,especially low-light images which are generally darker with low light and dark areas occupying the main part of the image,therefore,it is difficult to identify the details,which also brings serious challenges to the following image processing.Consequently,the corresponding image enhancement methods for processing such low-light images are employed,one of the typical methods is Retinex algorithm.The specific research content is as follows:（1）Illustrate the imaging environment and imaging characteristics of low-light images and the Retinex model commonly used to enhance low-light images,in addition,analyze some common algorithms based on Retinex theory.（2）In order to settle the problems of less details,color distortion and "halo" phenomenon,an improved Single-Scale Retinex algorithm based on Butterworth Low-pass Filter（BLPF）in HSV color space is proposed.The improved algorithm converts low-light images in different environments from RGB color space to HSV color space,on the one hand,the brightness component V is enhanced by improved Single-Scale Retinex based on BLPF for settling the problems of low brightness and less information in low-light images,on the other hand,the saturation component S is stretched adaptively in order to keep the original color of the image.Finally,the image is transferred back to RGB color space.The following testing results show that the improved algorithm can effectively reduce the noise interference,enhance the brightness of the dark area of the image,suppress the "halo" phenomenon of the local highlight area,highlight the details of the image,and restore the original color of the image.（3）Aiming at the problems that the traditional Retinex algorithm requires manual parameter adjustment and can only adapt to the limited low-light scenes,a low-light image enhancement algorithm based on convolutional neural network is proposed,which is named Retinex-Dn CNet,a data-driven network to learn low-light image decomposition and enhancement,which learns model parameters by end-to-end network training.The network is divided into two parts: Decom-Net and Enhance-Net.The Decom-Net decomposes the low-light image into reflectance（R）and illumination（I）.In order to solve the problems that the reflectance（R）has too much noise and the less details,the Enhance-Net improves the denoising convolutional neural network Dn CNN model firstly,and then uses the improved model New Dn CNN to denoise the reflectance（R）,while,aiming at the problem of low brightness and insufficient details of illumination（I）,the convolutional block attention model CBAM is introduced to enhance details and guide the network to correct the illumination.Finally,the adjusted illumination and reflectance are multiplied element by element for image reconstruction.The experimental results show that the enhanced low-light image brightness is improved,the details are prominent,the information is rich,the image distortion is small and the authenticity is close to nature.